KR20200046178A - Head region detection method and head region detection device - Google Patents

Abstract

The present invention relates to a head region detecting method which detects a human head in an input image to determine the number of people in the image, and to a head region detecting apparatus. According to an embodiment of the present invention, the head region detecting method comprises the steps of: receiving an initial image; setting a detection target region; generating a plurality of transformed images; detecting a first head region, which is a candidate estimated to have a head for each of the plurality of transformed images; detecting a second head region for each of the plurality of transformed images; and detecting a third head region which is a final head region in the detection target region.

Description

Head area detection method and head area detection device {HEAD REGION DETECTION METHOD AND HEAD REGION DETECTION DEVICE}

The present invention relates to a head region detection method and a head region detection apparatus, and more specifically, a head region detection method and a head region detection that detects a human head in an input image to determine how many people are in the image. It is about the device.

In recent years, people detection technology has been widely used in population analysis, traffic analysis, emergency alerts, games, etc. In order to collect various information or to create an environment for a certain space, it is required to efficiently calculate the number of people located in the space.

Previously, a method of counting the number of people included in an image captured by analyzing the image acquired by a depth camera or a stereo camera and detecting a human head in the image was used. In addition, recognition is performed by using a plurality of cameras to more accurately count the number of people, or by acquiring complex features such as color and depth of an image included in an image.

However, in the related art, the form that can be recognized by the human head is limited, and the recognition rate for the side or back of the head is reduced. Therefore, there is a problem in that a head of a person looking in a different direction without looking at the camera cannot be recognized. In addition, even if the entire body of a person is not included in the image or people overlap each other, there is a problem in that the recognition rate is lowered because the body is difficult to recognize.

Further, in order to improve the recognition rate of the head from the captured image, a depth camera may be used, or an image provided from a plurality of cameras photographing the same area may be used, but an additional cost is required compared to a general shooting environment. have.

In addition, the method based on the user's personal feature data such as movement or human body information must consider the differences of the individual face, height, and skin color together, so it is difficult to detect a person in various environments.

The head region detection method and the head region detection apparatus according to an embodiment of the present invention, in a method of detecting a human head from an input image, accurately recognize the head region where the head is located regardless of the state of the head represented on the image. The purpose is to detect quickly.

The head region detection method and the head region detection apparatus according to an embodiment of the present invention are performed without using an image input from a plurality of cameras or an image input through a depth camera, thereby reducing the cost of head detection. There is a purpose.

In order to solve the above problems, an object recognition method according to an embodiment of the present invention includes receiving an initial image; Setting an area to be detected in the initial image; Generating the detection target regions as a plurality of transformed images having different sizes; Detecting a first head region which is a candidate estimated to have a head for each of the plurality of transformed images; Analyzing the first head region and detecting a second head region for each of the plurality of transformed images; And analyzing the plurality of second head regions by overlapping the plurality of transformed images, thereby detecting a third head region that is a final head region in the detection target region. It may include.

The detection target area can be set by a user designating.

The generating of the plurality of transformed images may include selecting an object of the largest size and an object of the smallest size among objects located in the detection target area; And generating a converted image enlarged from the initial image or a reduced converted image from the initial image according to the size ratio of the largest object and the smallest object. It may include.

In the detecting of the first head region, through filtering on the transformed image, a feature map related to features of the region where the head is located is generated, and a convolutional neural network (CNN) operation for the generated feature map is set. Generating a first intermediate layer including a feature map of a compressed size by repeatedly performing the number of times; Repeating a CNN operation a predetermined number of times while maintaining the size of the feature map included in the first intermediate layer, and generating a plurality of second intermediate layers including the feature map according to each iteration step; And generating a third intermediate layer by performing a CNN operation by merging the feature maps of the second intermediate layers generated in the step of generating the plurality of second intermediate layers. It may include.

The step of generating the first head region may be a step of using a ReLU (Rectified Linear Unit) function in the process of generating intermediate layers.

The ReLU function may be a function having a negative slope in a negative range.

The second intermediate layer may include a larger number of feature maps than the first intermediate layer.

In the detecting of the second head region, a fourth intermediate layer is generated by repeatedly performing a CNN operation by merging the feature map of the second intermediate layer generated first among the second intermediate layers and the feature map of the third intermediate layer. To do; And generating a final layer by repeating the process of merging the feature map of the first intermediate layer to the generated feature map of the fourth intermediate layer. It may include.

The detecting of the third head region may include adjusting the plurality of transformed images of different sizes to have the same size; Comparing positions of second head regions included in the plurality of transformed images adjusted to the same size; And determining the third head region according to the degree of overlap of positions of the second head regions included in the plurality of transformed images. It may include.

The determining of the third head region according to the degree of overlap of the positions of the second head regions may be a step of determining the third head region through a non-maximum-suppression (NMS) algorithm.

An apparatus for detecting a head region according to an embodiment of the present invention for solving the above-described problem includes a memory in which at least one program is stored; And a processor operating under the control of the at least one program, wherein the processor receives an initial image and a detection target region, generates the detection target region as a plurality of transformed images having different sizes, and A first head region is detected as a candidate region where a head is estimated to be located for each of the plurality of transformed images, and the first head region is analyzed to detect a second head region for each of the plurality of transformed images, and the By analyzing a plurality of the second head regions by overlapping a plurality of transformed images, the third head region, which is the final head region in the detection target region, can be detected.

The detection target area can be set by a user designating.

The head region detection method and the head region detection apparatus according to an embodiment of the present invention, in a method of detecting a human head from an input image, accurately recognize the head region where the head is located regardless of the state of the head represented on the image. And can be detected quickly.

The head region detection method and the head region detection apparatus according to an embodiment of the present invention are performed without using an image input from a plurality of cameras or an image input through a depth camera, thereby reducing the cost of head detection. have.

In addition, the process of generating the background image in the process of detecting the head region is not required, and it is not necessary to store the image or feature information for a specific operation or environment, so it has an excellent usability effect even in a mobile terminal having relatively low computational processing performance. have.

However, the effects that can be achieved according to an embodiment of the present invention are not limited to those mentioned above, and other effects not mentioned are those skilled in the art from the following description. Will be clearly understood.

1 is a view for explaining a head region detection method according to an embodiment of the present invention.
FIG. 2 is a diagram for explaining in more detail the steps of generating a plurality of transformed images in the embodiment of FIG. 1.
3 is a diagram for explaining an example of a process in which an object is selected in a detection target area.
4 is a view for explaining that a plurality of transformed images are generated.
FIG. 5 is a diagram for explaining the first head region detection step in more detail in the embodiment of FIG. 1.
6 is another diagram for describing a first head region detection step.
7 is another diagram for describing a first head region detection step.
8 is a diagram for explaining a sub-sampling process of the first head region detection step by way of example.
9 is a view for explaining a graph of the ReLU function.
FIG. 10 is a view for explaining the second head region detection step of FIG. 1 in more detail.
11 is another diagram for describing a second head region detection step.
12 is another diagram for describing a second head region detection step.
13 is a view for explaining the third head region detection step of FIG. 1 in more detail.
14 is a view for explaining a head region detection apparatus according to another embodiment of the present invention.

The present invention can be applied to various changes and can have various embodiments, and specific embodiments are illustrated in the drawings and will be described in detail through detailed description. However, this is not intended to limit the present invention to specific embodiments, and it should be understood that the present invention includes all modifications, equivalents, and substitutes included in the spirit and scope of the present invention.

In describing the present invention, when it is determined that a detailed description of related known technologies may unnecessarily obscure the subject matter of the present invention, the detailed description will be omitted. In addition, the numbers (for example, first, second, etc.) used in the description process of the present specification are only identification numbers for distinguishing one component from other components.

Further, in this specification, when one component is referred to as "connected" or "connected" with another component, the one component may be directly connected to the other component, or may be directly connected, but in particular, It should be understood that, as long as there is no objection to the contrary, it may or may be connected via another component in the middle.

1 is a view for explaining a head region detection method according to an embodiment of the present invention. Referring to FIG. 1, the head region detection method includes an initial image input step (S100), a detection target region setting step (S200), a plurality of transformed image generation steps (S300), a first head region detection step (S400), and a second. A head region detection step (S500) and a third head region detection step (S600) may be included.

Each step of the head region detection method according to an embodiment of the present invention may be performed by a processor of the head region detection apparatus, which will be described later. The processor may load and use data necessary for performing the above steps from the memory of the head region detection device, and process an operation necessary for performing the above steps.

The initial image input step (S100) may mean a step in which the processor of the head region detection apparatus loads an initial image from an internal or external database or receives an initial image transmitted directly from a photographing unit of the head region detection apparatus. The initial image may be image data for a still picture or picture, and the image data may mean one of frames of an image in which a plurality of frames are displayed according to a sequence of sequences.

The detection target region setting step S200 may mean a step of setting all or part of the initial image as a detection target region. More specifically, the detection target region may be set by a user of the head region detection device designating an area in which the head region needs to be detected in the initial image. The detection target area may be set for the entire initial image, if necessary.

The generating of the plurality of transformed images (S300) may mean a step in which the head region detecting apparatus generates a region to be detected as a plurality of transformed images having different sizes.

Here, the converted image may be a plurality of images generated by expanding or reducing the width and length of the detection target area at the same ratio. That is, the detection target region of the initial image may be generated with at least two or more transformed images, and the plurality of transformed images may also include an image of the detection target region whose ratio is not changed. The plurality of transformed images may be subsequently performed through the first to third head region detection steps. The details of the plurality of transformed image generation steps S300 will be described later.

The first head region detection step (S400) may mean a step in which the head region detection apparatus detects a first head region, which is a candidate region estimated to be located with respect to each of the plurality of transformed images.

Here, the first head region refers to a region in which a human head is estimated to be located for each of a plurality of transformed images generated in different sizes. The first head region may mean a head region that is primarily detected through a basic operation prior to the second and third head regions that are subsequently detected. That is, the first head region is an area used for detection of the second head region and the third head region. The head region detecting apparatus sequentially detects from the first head region to the second head region and the third head region, and increases the probability that a human head exists in the head region.

When the first head region detection step S400 is performed, the first head region may be detected for each of the plurality of transformed images. Since the transformed images have different sizes from each other, the first head region detected in each transformed image may not be the same. Details of the first head region detection step S400 will be described later.

The second head region detection step S500 may mean a step in which the head region detection apparatus analyzes the first head region and detects the second head region for each of the plurality of transformed images.

The second head region detection step S500 may be a step of increasing the probability of detecting the head region for the region in which the characteristics of the human head are more strongly included, except for the head region that is incorrectly detected in the first head region. As a result, by performing the second head region detection step S500 for each transformed image, it is possible to detect the second head region having a higher probability that the human head is located than the first head region.

Details of the second head region detection step S500 will be described later.

In the third head region detection step (S600), the head region detection apparatus analyzes a plurality of second head regions by overlapping a plurality of transformed images, thereby detecting a third head region as a final head region in the detection target region. You can.

The head region detection apparatus may generate a second head region detected for each converted image through the preceding second head region detection step (S500). The head region detection apparatus may re-adjust each converted image to the same size in the third head region detection step (S500). The head region detecting apparatus may determine the degree of overlap of the second head region between the re-adjusted transformed images, and finally detect the third head region estimated to be located at the human head.

The detailed description of the third head region detection step S600 will be described later.

The first head region detection step (S400) and the second head region detection step (S500) may be performed for each of the plurality of transformed images generated in the plurality of transformed image generating stages (S300). At this time, after the first head region detection step (S400) and the second head region detection step (S500) are sequentially performed on one transformed image, the first head region detection step (S400) is performed on the transformed image having different sizes. And the second head region detection step S500 may be repeatedly performed sequentially. Also, in another embodiment, after the first head region detection step S400 is performed on each of the plurality of transformed images, the second head region detection step S500 may be performed on each of the plurality of transformed images again. There is also.

The first head region detection step S400 and the second head region detection step S500 may be performed using a convolutional neural network (CNN) operation. In more detail, when the first head region detection step (S400) and the second head region detection step (S500) are performed, the feature map included in each intermediate layer proceeds to the intermediate layer of the next process, and performs CNN calculation. Through this, a feature map with enhanced features of the head region may be output.

2 is a view for explaining in more detail the step (S300) of generating a plurality of transformed images in the embodiment of FIG. 1. Referring to FIG. 2, a plurality of transformed image generation steps (S300) may include an object selection step (S310) and an enlarged or reduced transformed image generation step (S320).

The object selection step S310 may mean a step of selecting an object having the largest size and an object having the smallest size among objects located in the detection target area.

The head region detection apparatus may select an arbitrary object, not a background, within the detection target region, and may select an object of the largest size and an object of the smallest size among objects of a kind similar to the selected object. In an embodiment of the present invention, the object may mean the body of a person, including the head. The object selection step S310 may be a step of selecting a target object through simple and fast calculation and selecting a large-sized object and a small-sized object prior to the processing of the specific operation.

In the step of generating an enlarged or reduced transformed image (S320), a transformed image enlarged from the initial image is generated according to a size ratio of the largest object and the smallest object, or the converted image reduced from the initial image is generated. It can mean the step of creating.

When a plurality of transformed images are generated, the head region detecting apparatus selects the object selection step (S310) in order to determine a ratio of the transformed image to be enlarged or reduced, or how many images to be generated. Results are available. That is, the head region detecting apparatus may determine the ratio of the plurality of transformed images and the number of transformed images generated by comparing the size of the largest object selected in the object selection step S310 and the size of the smallest sized object.

In more detail, the converted image may be changed according to the size of the set detection target region and the size of the selected object in the detection target region. For example, when the size of the detection target area is N and the size of the smallest selected object is A, the conversion ratio Rmin of the converted image for the object may be set to A / N. Also, when the object size of the largest selected size is B, the ratio Rmax of the converted image for the object may be set to B / N. In addition, it is assumed that a ratio R1 value, which is a reference for generating a converted image set in advance or arbitrarily set by a user, is set to C / N, Rmin is 0.5 times R1, and Rmax is 2 times R1. At this time, the head region detection apparatus may generate a transform image in which the size of the detection target region is doubled based on the smallest size object and a transform image reduced to 0.5 based on the largest size object. In the above example, the head region detection apparatus exemplifies generating two transformed images of which the size is increased or decreased, but the transformed images generated according to the setting may be three or more, and the enlarged or reduced ratio may be varied. .

3 is a diagram for explaining an example of a process in which an object is selected in a detection target area. Referring to FIG. 3, it is possible to check a state in which the detection target area B is set by the user in the initial image A. The head region detecting apparatus may recognize the entire human body as an object in the detection target region B, and may select the largest object C and the smallest object D from among the recognized objects. In FIG. 3, the object recognized by the head region detection device as an object is illustrated as a human body, but this may be changed according to a user's or designer's setting or environment.

4 is a view for explaining that a plurality of transformed images are generated. Referring to FIG. 4, when an area to be detected of an initial image is selected and an object according to a size is selected in the area to be detected, it can be seen that a plurality of transformed images are generated corresponding to the size of the selected object. As illustrated in FIG. 4, the size of the transformed images may be different, and the transformed image S generated by being enlarged to the largest size may increase the probability of detecting a small sized head region. Conversely, the transformed image L generated by being reduced to the smallest size has an effect of increasing the probability of detecting a large sized head region.

FIG. 5 is a diagram for explaining the first head region detection step in more detail in the embodiment of FIG. 1. Referring to FIG. 5, the first head region detection step (S400) includes generating a first intermediate layer (S410), generating a second intermediate layer (S420), and generating a third intermediate layer (S43). ).

In operation S410 of generating a first intermediate layer, a feature map related to a feature of a region in which the head is located is filtered through a transformation image, and a CNN (Convolutional Neural Network) operation on the generated feature map is performed. By repeatedly performing the set number of times, it may mean a step of generating a first intermediate layer including a feature map of a compressed size.

The head region detecting apparatus may repeatedly perform CNN calculation for each of the plurality of transformed images generated in the previous step. The converted image may be output as a plurality of feature maps while the first head region detection step S400 is performed. The CNN operation is repeatedly performed on one transformed image and a layer may be generated, and such a layer may include a plurality of feature maps differentiated into elements for detecting a head region in the transformed image. The first intermediate layer may be understood as a set of layers including the same number of feature maps.

In addition, the first intermediate layer generation step (S410) may be a step in which convolution and subsampling of the CNN are repeatedly performed on the transformed image, and the feature map size of the layer to be analyzed is reduced. .

The number of times the CNN is repeatedly performed in the first intermediate layer generation step S410 may vary according to the designer's intention of the head region detection method. In general, as the CNN operation is repeated, the detection accuracy of the head region increases, but a problem in which the calculation time is increased may occur.

In the first intermediate layer generation step (S410), the first intermediate layer of the immediately preceding step may be used to generate the first intermediate layer of the next step.

In operation S420 of generating a second intermediate layer, the CNN operation is repeatedly performed a predetermined number of times while maintaining the size of the feature map included in the first intermediate layer, and includes a feature map according to each iteration step. It may mean a step of generating a plurality of second intermediate layers. Here, the second intermediate layer may include a larger number of feature maps than the first intermediate layer.

The second intermediate layer generation step S420 may be understood as a step in which an additional CNN operation is performed on the first intermediate layer generated last through the first intermediate layer generation step S410. In the second intermediate layer generation step (S420), the size reduction of the feature map through subsampling is not performed. This is because the second intermediate layer generation step S420 includes a greater number of feature maps than the first intermediate layer.

That is, the feature map of the second intermediate layer generated in the second intermediate layer generation step S420 may include a larger number of feature maps than the first intermediate layer and may have a smaller size. The second intermediate layer generation step (S420) generates a layer including a larger number of feature maps than the first intermediate layer generation step (S410), but generates fewer feature maps, but has fewer operations. Various features can be used to improve the accuracy of head region detection.

The step of generating a third intermediate layer (S430) means generating a third intermediate layer by performing a CNN operation by merging the feature maps of the plurality of second intermediate layers and the feature maps of the first intermediate layer. You can.

The third intermediate layer may refer to a layer finally generated through the CNN operation of the first head region detection step (S400). The third intermediate layer may be generated by merging and calculating not only the second intermediate layer finally calculated and generated, but also a plurality of second intermediate layers generated in the previous step. In generating the third intermediate layer, a variety of feature maps can be used by using the preceding intermediate layers together, thereby increasing the probability of head region detection.

Here, the head region detection device may use a ReLU (Rectified Linear Unit) function in the process of generating the first to third intermediate layers. The ReLU function is a function that assumes a negative number of 0 to remove noise generated in the process of generating the intermediate layers.

And, in another embodiment, ReLU may be a function having a negative slope in a negative range. In the process of generating an intermediate layer in the method of ReLU used in the above-described embodiment, in order to remove noise and improve the detection accuracy of the head region, it is possible to have a predetermined slope without fixing a negative number to zero. At this time, the value of the weight used for the slope is not derived in the learning process, but by using a fixed value in advance, there is an effect of reducing the amount of computation required in the learning process and allowing computation at a higher speed.

6 is another diagram for describing a first head region detection step.

Referring to FIG. 6, a plurality of CNN operations (Conv) are performed on the converted image, generating a first intermediate layer (S410), generating a second intermediate layer (S420), and a third intermediate layer. It can be seen that the step of generating (S430) is illustratively illustrated.

In FIG. 6, in step S410 of generating the first intermediate layer, it can be seen that 6 CNN operations are repeatedly performed on the input transformed image. During the operation (S410) of generating the first intermediate layer, six first intermediate layers (Layer 1 to Layer 6) may be generated. The first first intermediate layer (Layer 1) is generated based on the transformed image, and each subsequent first intermediate layer (Layer 2 to Layer 6) may be generated based on the first intermediate layer of the previous step.

Thereafter, a step S420 of generating a second intermediate layer is performed. At this time, the first generated first intermediate layer (Layer 6) may be used at the start of the second intermediate layer generation step (S420), it can be seen that the 4 CNN operation is repeatedly performed. While the second intermediate layer generation step (S420) is performed, four second intermediate layers (Layer 7 to Layer 10) may be generated. The first second intermediate layer (Layer 7) is generated based on the first generated first intermediate layer (Layer 7), and each subsequent second intermediate layer (Layer 8 to Layer 10) is the second intermediate layer of the previous step. It can be generated on the basis of.

Finally, the third intermediate layer generation step (S430) may be generated by merging all of the second intermediate layers (Layer 7 to Layer 10). Thereafter, an output (out) of the first head region detection step may be generated through the third intermediate layer (Layer 11). As described above, by merging and processing the second intermediate layers (Layer 7 to Layer 10) previously generated together, a feature map indicating recognition and location information for various shapes and sizes of the head region can be more accurately extracted. You can. Therefore, there is an effect of improving the detection probability even for hair regions of various sizes and shapes.

In FIG. 6, six first intermediate layers (Layer 1 to Layer 6) and four second intermediate layers (Layer 7 to Layer 10) are illustrated, but these are merely numerical values set for convenience of description. Therefore, the number may vary depending on the intention of the designer or the user of the head region detection method.

7 is another diagram for describing a first head region detection step.

Referring to FIG. 7, the contents of each intermediate layer generated in the process of performing the first head region detection step described in FIG. 6 can be seen in more detail. In FIG. 7, CONV is a convolution, and RELU is a ReLU (Rectified Linear Unit) function. Ker is the size of the kernel, pad is the size of the zero padding. After the CNN operation is performed through zero padding, an edge pixel value for which the operation is not performed is assigned to 0 to preserve the size of the resulting image.

In performing CNN operation for each layer, a RELU function may be used, and in this case, a weight for a negative region may be set to 0.01 (negative slope). Depending on the layer, the RELU function may or may not be applied. In the application of the RELU function, the weight used or whether to use the RELU function for each intermediate layer may vary depending on the designer's intention of the head region detection method.

In FIG. 7, it is illustrated that the first intermediate layers (Layers 1 to 6) output A feature maps, and the second intermediate layers (Layers 7 to 10) output B feature maps. The first intermediate layers (Layers 1 to 6) repeatedly utilize A feature maps to efficiently extract the input image and features extracted in the intermediate stage. The number of feature maps generated in the second intermediate layers (Layers 7 to 10) is at least twice that of the first intermediate layers (Layers 1 to 6), compared to the first intermediate layers (Layers 1 to 6). The processing speed can be increased by using feature maps extracted from more filters while being relatively small in size. In addition, since the above-described sub-sampling step is not performed on the feature map, it is possible to efficiently utilize a threshold value to increase the efficiency of the operation.

A feature map corresponding to N pieces of information to be detected is derived from the third intermediate layer (Layer 11). For example, when only the head region is to be detected, one feature map may be derived to detect one object. In addition, when it is desired to secure a detection area for a portion other than the head area, the value of N may be set to a number of 2 or more.

8 is a diagram for explaining a sub-sampling process of the first head region detection step by way of example.

Referring to FIG. 8, in performing the first intermediate layer generation step (S410) on the transformed image, a process of convolution and subsampling operations can be seen. As shown in FIG. 8, after forming a first intermediate layer by using a patch having a size of 5x5 pixels, a sub-sampling operation is performed by performing operations such as maximum and average values in the 2x2 patch to obtain a feature map having a small size.

9 is a view for explaining a graph of the ReLU function.

FIG. 9 is for explaining a rectified linear unit (ReLU) and a Leaky ReLU used in the first intermediate layer generation step (S410) and the second intermediate layer generation step (S420). When reducing noise in the middle layer, a ReLU function that assumes a negative number as 0 can be used. In addition, in another embodiment, a fixed weight (ai) may be applied to a negative value in order to reduce the noise of the intermediate layer and at the same time increase the accuracy, and utilize it. At this time, by using a fixed value whose weight ai is not derived from the learning process, learning can be performed quickly and the number of data to be stored for calculation can be reduced.

FIG. 10 is a view for explaining the second head region detection step of FIG. 1 in more detail.

Referring to FIG. 10, the second head region detection step (S500) may include a fourth intermediate layer generation step (S510) and a final layer generation step (S520).

The fourth intermediate layer generation step (S510) generates a fourth intermediate layer by repeatedly performing a CNN operation by merging the feature map of the second intermediate layer first generated among the second intermediate layers and the feature map of the third intermediate layer. It can mean a step.

The final layer generation step S520 may refer to a step of generating a final layer by repeating the process of merging the feature map of the first intermediate layer to the generated feature map of the fourth intermediate layer.

The fourth intermediate layer generation step (S510) and the final layer generation step (S520) will be described with reference to FIGS. 11 and 12.

11 is another diagram for describing a second head region detection step. The examples of FIGS. 11 and 12 will be described again using the examples of FIGS. 6 and 7 described above.

In FIG. 11, the second intermediate layer (Layer 7) may refer to the second intermediate layer (Layer 7) that is first generated in the previous second intermediate layer generation step (S400). That is, the second intermediate layer (Layer 7) may refer to an intermediate layer in which the CNN operation for the first intermediate layer (Layer 6) last generated through the first intermediate layer generation step (S300) is processed.

In FIG. 11, the fourth intermediate layer merges the output (out) and the second intermediate layer (Layer 7) generated as a result of the calculation of the third intermediate layer generation step (S430) to form the first fourth intermediate layer 710. Can be created. Thereafter, the second intermediate layer (Layer 7) is merged with the fourth intermediate layer 710 to generate the fourth intermediate layer 720 in the next step.

That is, the first second intermediate layer generated in the previous process may be used for merging in the process of generating the fourth intermediate layer and generating the fourth intermediate layer in the next step. The number of steps of repeating the CNN operation for the generated fourth intermediate layer may be differently set according to the intention of the designer of the head region detection method. As the CNN operation for the fourth intermediate layer is repeated, it is possible to further improve the probability of detecting where the head region is located in the detection target region.

The head region detection device can repeat this process a predetermined number of times to derive accurate results while maintaining differences and discrimination between features. In other words, the final accuracy can be increased by excluding the false-detected candidate region and increasing the probability of the part to be detected.

When the fourth intermediate layer generation step S510 ends, a final layer generation step S520 of generating a final layer for the converted image may be performed.

The final layer means the layer finally determined as the head region located in the converted image. A region estimated as a head region according to the final layer may be defined as a second head region.

12 is another diagram for describing a second head region detection step.

In FIG. 12, CONV is a convolution, and RELU is a ReLU (Rectified Linear Unit) function. Ker is the size of the kernel, pad is the size of the zero padding. After the CNN operation is performed through zero padding, an edge pixel value for which the operation is not performed is assigned to 0 to preserve the size of the resulting image.

In FIG. 12, the fourth intermediate layers (Layers 21 to 24) are illustrated as outputting as many as C feature maps. The fourth intermediate layers (Layers 21 to 24) repeatedly utilize C feature maps to efficiently extract the input image and features extracted in the intermediate stage.

The final layer (Layer 25) is generated through the CNN operation for the last generated fourth intermediate layer (Layer 24), and may include T final feature maps. Here, T may mean one of the parts of the body to be detected, and the T value is only 1 for detecting only the head region. At this time, the final layer (Layer 25) may be generated as many as the number of transformed images.

13 is a view for explaining the third head region detection step of FIG. 1 in more detail. Referring to FIG. 13, the third head region detection step (S600) may include a transformed image size adjustment step (S610), a second head region position comparison step (S620), and a third head region determination step (S630). .

The converted image size adjustment step S610 may mean adjusting the plurality of converted images of different sizes to the same size.

In the previous step, the final layer is obtained as many as the number of transformed images. At this time, since each final layer has a different size of the feature maps included with each other in correspondence to the transformed images of different sizes, it is necessary to re-adjust to a reference size. Accordingly, the transform image size adjustment step S610 may be a step of resizing all of the plurality of transformed images to the same size, and converting the size of the final layers corresponding to the transformed images to be the same.

The second head region position comparison step S620 may mean comparing the positions of the second head regions included in the plurality of transformed images adjusted to the same size.

The determining of the third head region (S630) may mean determining the third head region according to the degree of overlap of positions of the second head regions included in the plurality of transformed images. In this case, the determining of the third head region (S630) may determine the third head region through a non-maximum-suppression (NMS) algorithm. Here, the third head region refers to a region where the head region detection device finally determines that it is the head region within the detection target region.

More specifically, resized transformed images can be positioned on the same area. At this time, the third head region may be detected by performing a non-maximum-suppression (NMS) according to the degree of overlap (IOU, intersection over union) in consideration of the block size of the feature map of each second head region. That is, when the degree of overlap between the second head regions is greater than or equal to a predetermined value, the head region detection apparatus may determine the corresponding region as the third head region. In this case, the block size may utilize the size and correction threshold of the feature map compared to the reference image. For example, if the size of the reference image is 400x400 and the size of the feature map is 50x50, the block size may be set to 400 / 50x4.

14 is a view for explaining a head region detection apparatus according to another embodiment of the present invention. Referring to FIG. 14, the head region detection apparatus 100 may include a memory 110, a processor 120, and a photographing unit 130. In describing FIG. 14, description of the same configuration or effect as the above-described drawings will be omitted.

The head area detection device 100 is a device for recognizing and detecting a target object from an input image, and operates the overall service such as configuring a service screen, data input, data transmission, and data storage under the control of a web / mobile site or a dedicated application. You can do The head region detection device 100 includes a photographing unit 130, and includes a variety of devices such as a desktop computer, a notebook computer, a PDA, a web pad, and a tablet PC, as well as a smartphone capable of graphic processing from an input image. can do.

At least one program may be stored in the memory 110. In addition, files related to the first to third models to be described later may be stored, and stored data may be provided according to a request of the processor.

The processor 120 may process an operation for recognizing a target object from input image data.

The photographing unit 130 may photograph a target object. The photographing unit 130 may include a camera embedded in the object recognition device or an external camera connected to a separate device, and may generate image data by photographing a target object in the form of a photo or video.

In more detail, the processor 120 receives an initial image and a detection target region, generates the detection target region as a plurality of transformed images having different sizes, and a head is positioned for each of the plurality of transformed images. A first head region is detected as an estimated candidate region, and the first head region is analyzed to detect a second head region for each of the plurality of transformed images, and the plurality of transformed images are superimposed to overlap the plurality of second regions. By analyzing the head region, the third head region, which is the final head region in the detection target region, can be detected.

The?, The detection target area can be set by designating the user of the head area detection device.

The embodiments of the present invention have been described above with reference to the accompanying drawings, but those of ordinary skill in the art to which the present invention pertains can implement the present invention in other specific forms without changing its technical spirit or essential features. You will understand that there is. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive.

S100: Initial image input step
S200: detection target area setting step
S300: multiple conversion image generation step
S400: first head area detection step
S500: second head area detection step
S600: Third head area detection step

Claims (12)

Receiving an initial image;
Setting an area to be detected in the initial image;
Generating the detection target regions as a plurality of transformed images having different sizes;
Detecting a first head region which is a candidate estimated to have a head for each of the plurality of transformed images;
Analyzing the first head region and detecting a second head region for each of the plurality of transformed images; And
Detecting a third head region as a final head region in the detection target region by analyzing the plurality of second head regions by overlapping the plurality of transformed images; Head area detection method comprising a. The method of claim 1, wherein the detection target region
A head area detection method characterized by being set by a user designating. The method of claim 1, wherein the generating of the plurality of transformed images
Selecting an object of the largest size and an object of the smallest size among the objects located in the detection target area; And
Generating a converted image enlarged from the initial image or a reduced converted image from the initial image according to a size ratio of the largest object and the smallest object; Head area detection method comprising a. The method of claim 1, wherein detecting the first head region
After filtering the transformed image, a feature map related to the feature of the region where the head is located is generated, and CNN (Convolutional Neural Network) operation on the generated feature map is repeatedly performed a predetermined number of times to obtain a feature of the compressed size. Creating a first intermediate layer including a map;
Repeating a CNN operation a predetermined number of times while maintaining the size of the feature map included in the first intermediate layer, and generating a plurality of second intermediate layers including the feature map according to each iteration step; And
Generating a third intermediate layer by performing a CNN operation by merging feature maps of the second intermediate layers generated in the generating of the plurality of second intermediate layers; Head area detection method comprising a. The method of claim 4, wherein generating the first head region
A method of detecting a head region, characterized in that the step of using a ReLU (Rectified Linear Unit) function in the process of generating the intermediate layers. The method of claim 5, wherein the ReLU function
A head region detection method characterized in that it is a function having a negative slope in a range of negative numbers. The method of claim 4, wherein the second intermediate layer
A head area detection method comprising a feature map having a larger number of feature maps than the first intermediate layer. The method of claim 4, wherein detecting the second head region
Generating a fourth intermediate layer by repeatedly performing a CNN operation by merging the feature map of the second intermediate layer first generated from the second intermediate layer and the feature map of the third intermediate layer; And
Generating a final layer by repeating the process of merging the feature maps of the first intermediate layer to the generated feature maps of the fourth intermediate layer; Head area detection method comprising a. The method of claim 1, wherein detecting the third head region
Adjusting the plurality of transformed images of different sizes to the same size;
Comparing positions of second head regions included in the plurality of transformed images adjusted to the same size; And
Determining the third head region according to a degree of overlap of positions of the second head regions included in the plurality of transformed images; Head area detection method comprising a. 10. The method of claim 9, The step of determining the third head region according to the degree of overlap of the position of the second head region
And determining the third head region through a NMS (Non-Maximum-Suppression) algorithm. A memory in which at least one program is stored; And
It includes; a processor that operates under the control of the at least one program; includes,
The processor
An initial image and a detection target region are inputted, the detection target region is generated as a plurality of transformed images having different sizes, and a first head region is a candidate region estimated to have a head for each of the plurality of transformed images And detecting the second head region for each of the plurality of transformed images by analyzing the first head region, and analyzing the plurality of second head regions by overlapping the plurality of transformed images, thereby detecting the target A head region detection device characterized in that it detects a third head region which is the final head region in the region. The method of claim 11, wherein the detection target area
A head area detection device characterized by being set by a user designating.

Images